A lightweight recursive attention-based multi-channel data recovery framework for structural health monitoring

In wireless sensor networks, the measurement transmission from sensor nodes to a data center is prone to data loss due to unstable communication and limited bandwidth, thereby compromising the reliability of subsequent structural health monitoring (SHM) tasks. Recent progress in deep learning has provided effective solutions for data recovery, but most existing approaches rely on computationally intensive architectures, limiting their use in resource-constrained monitoring environments. This study proposes a lightweight recursive attention-based framework for efficient multi-channel data recovery. The design integrates the windowing operation and recursive attention mechanism to reduce the complexity of attention calculations while preserving temporal dependencies across sensor measurements. The framework is validated using two real-world SHM datasets, including vibration tests on a reduced-scale curved bridge model and field monitoring data from the Canton Tower. Results demonstrate that the proposed method achieves recovery accuracy comparable to or better than existing models, with substantially lower computational demand. Owing to its lightweight architecture and verified edge deployment capability, the framework supports real-time data recovery in wireless SHM systems, thereby enhancing the robustness and reliability of SHM deployment.